This study proposes a machine learning model based on recurrent neural networks (RNN) for the early detection of late onset sepsis (LOS) in premature infants, using heart rate variability (HRV) as input. The proposed model was designed to be suitable for use as a decision support system (DSS) in neonatal intensive care units (NICU).
Methods: HRV data was acquired from 259 infants born prematurely. Infants in the population were retrospectively classified into a control group (n = 218), and a group of infants who developed LOS (n = 41). The HRV features were used in a machine learning model based on RNN architecture, which combines gated recurrent units and long short-term memory units. The output of the model is the probability of the patient having LOS, which is calculated in time-steps of 30 minutes. To train and test the model, the control and LOS groups were split into training and testing set (75\% and 25\% of each group, respectively). The performance of the model was measured using the area under the receiver operating characteristics curve (AUROC) as main metric.
Results: The model achieved an AUROC of more than 80% for the 24 hours before the onset of LOS, reaching a maximum of 90.4% (95\% CI [88.1%, 92.6%]) six hours before the time of infection onset.
Conclusion: This method has the potential to be easily implemented as a DSS for real-time LOS detection in NICU, as it only uses data which is continuously available in such settings, and produces an updated probability of LOS every 30 minutes. The use of a RNN model for this task allows to account for changes over time, induced by the infection onset, without the need for feature engineering, while achieving an AUROC higher than 80% 24 hours before sepsis onset.